1997-12-15 [should be 1997-12-25 ?] egcs snapshot available -m68k-next-nextstep3 results

[Toon Moene]

I successfully built egcs-971225 on m68k-next-nextstep3.

Running c-torture-1.45 on it (dejagnu doesn't build on this  
machine) I got one `regression':

% diff egcs-971225.cto?t egcs-1.0.cto?t
321,328d320
< ERROR: 961203-1.c: compiler returns exit status 1:
< ERROR: 961203-1.c: compiler returns exit status 1: -O  
-fno-omit-frame-pointer
< ERROR: 961203-1.c: compiler returns exit status 1: -O  
-fomit-frame-pointer
< ERROR: 961203-1.c: compiler returns exit status 1: -O2  
-fno-omit-frame-pointer
< ERROR: 961203-1.c: compiler returns exit status 1: -O2  
-fomit-frame-pointer
< ERROR: 961203-1.c: compiler returns exit status 1: -O3  
-fno-omit-frame-pointer
< ERROR: 961203-1.c: compiler returns exit status 1: -O3  
-fomit-frame-pointer
< ERROR: 961203-1.c: compiler returns exit status 1: -O3  
-funroll-all-loops
561c553
< Test completed with 32 failures
---
> Test completed with 24 failures

This isn't a real regression, however:

% /usr/test/bin/gcc c-torture-1.45/compile/961203-1.c
c-torture-1.45/compile/961203-1.c:2: size of array `a' is too large

Alias analysis did make a step backwards, though.  Here's a routine  
I used about a year ago on the gcc2 mailing list to show that  
John's alias analysis really works:

      subroutine MXMA( A, IA, JA, B, IB, JB, C, IC, JC, M, K, N )
      implicit none
      integer  i, j, l, ia, ja, ib, jb, ic, jc, m, k, n
      double precision   a(*), b(*), c(*)

      if ( m  .lt. 0 .or.
     ,     n  .lt. 0 .or.
     ,     k  .lt. 0 .or.
     ,     ia .eq. 0 .or.
     ,     ja .eq. 0 .or.
     ,     ib .eq. 0 .or.
     ,     jb .eq. 0 .or.
     ,     ic .eq. 0 .or.
     ,     jc .eq. 0     ) then
         write(0,*)'MXMA - ARGUMENT ERROR; ABORT'
         call abort
      endif

      if ( m .eq. 0 .or. n .eq. 0) return

      do i = 1, m
         do j = 1, n
            c((i-1)*ic+(j-1)*jc+1) = 0.0
         enddo
      enddo

      if ( k .eq. 0 ) return

      do i = 1, m
         do j = 1, n
            do l = 1, k
               c((i-1)*ic+(j-1)*jc+1) = c((i-1)*ic+(j-1)*jc+1) +
     ,            a((i-1)*ia+(l-1)*ja+1) * b((l-1)*ib+(j-1)*jb+1)
            enddo
         enddo
      enddo

      return
      end

This is the inner loop as produced by previous versions of egcs:

L28:
        movel a6@(8),a1
        fmoved a1@(d0:l:8),fp0
        movel a6@(20),a1
        fmuld a1@(d1:l:8),fp0
        faddd a0@,fp0
        fmoved fp0,a0@
        addl a6@(-4),d1
        addl d6,d0
        dbra d2,L28

And this is what we get now:

L28:
        movel a2,d3
        mulsl a5@,d3
        movel a0,d0
        mulsl a4@,d0
        addl d0,d3
        movel a2,d2
        movel a6@(12),a3
        mulsl a3@,d2
        movel d5,d0
        movel a6@(16),a3
        mulsl a3@,d0
        addl d0,d2
        movel d5,d1
        movel a6@(24),a3
        mulsl a3@,d1
        movel a0,d0
        movel a6@(28),a3
        mulsl a3@,d0
        addl d0,d1
        movel a6@(8),a3
        fmoved a3@(d2:l:8),fp0
        movel a6@(20),a3
        fmuld a3@(d1:l:8),fp0
        movel a6@(32),a3
        faddd a3@(d3:l:8),fp0
        fmoved fp0,a3@(d3:l:8)
        addql #1,d5
        dbra d4,L28

The abundance of integer multiplies clearly shows that strength  
reduction couldn't find the induction variables, most probably  
because alias analysis didn't proclaim ia, ib, ic, ja, jb, and jc  
loop invariants (note: g77 passes by reference, so alias analysis  
must prove that c doesn't overlap these scalar variables !).

Toon.